Today I tested Phi-3-Mini-4k-Instruct-q4-GGUF (3.2B) in my Asus K541UJ 
with 8GB RAM, an i3 and a GeForce 920M (2GB or 4GB VRAM I think) in Jan
(its like LLM studio but OpenSource, its a bad worse since for most 
of the models you have to upload it yourself, but its OpenSource and
LLMStudio is freeware and propietary so I don't believe them) 
And the models was acceptable, it generated a robotic story and almost
got right the first declension in latin (it failed in the ablative sg)
and refused to generate smut or tell me the dick size of early medieval
knights.
Maybe it was because it was a quantised or maybe I didn't know what to
ask, maybe I test it again with homework but for what I tested today
I give it a 6.5/10, it is better than TinyLlama (1.1B), I don't think
I can run something bigger than 3.2B, maybe a q3 of Llama 3 8b?